feat: Korpus-Kuratierungsrichtlinie - Floors und Leitplanke fuer reaktive Fixes (4.2.0, #28)

Files changed:
- CHANGES.md
- VERSION
- instructions/dev/corpus-policy.md
- instructions/dev/stack-dev/SKILL.md

CHANGES.md

@@ -20,6 +20,142 @@ their date-only headings.
 
 ---
 
+## 4.2.0 - 2026-09-03 - Korpus-Kuratierungsrichtlinie: Untergrenzen und Leitplanke für reaktive Fixes
+
+**Author:** Torben Nehmer
+
+Ein Demo-Korpus will klein und stabil sein, ein Testbett groß, unordentlich und in Bewegung -
+dieses Repo verlangt seit der Veröffentlichung beides vom selben `kb/` (Gitea #28). Die Sitzung
+vom 2026-09-02 hatte Fixture, `--with-demo` und ein zweites Repo bereits verworfen; offen blieb
+nur, wie kuratiert "kuratiert genug" heißt und welche Leitplanke reaktive Fixes bekommen.
+
+**Neu:** `instructions/dev/corpus-policy.md`. Fünf Untergrenzen, jede mit einer bestehenden
+`wikitool`-Prüfung messbar, keine davon durch neuen Tool-Code: jeder Seitentyp und jeder
+deklarierte Subtyp mit mindestens einer Seite, mindestens fünf Seiten mit mindestens drei
+Quellen, ein bis zehn Orphan-Seiten, im Schnitt mindestens vier ausgehende Wikilinks pro Seite.
+Gemessen am 2026-09-03: 181 Seiten, alle Typ-/Subtyp-Floors erfüllt, 12 Seiten mit ≥3 Quellen, 3
+Orphans, Ø 6,2 ausgehende Links - der Korpus war bereits groß genug, ohne dass eine einzige
+Seite eigens dafür angelegt werden musste. Eine Untergrenze wird nie durch eine erfundene Seite
+gefüllt, sondern durch eine echte Quelle beim nächsten passenden Ingest - Invariante 3 gilt
+unverändert.
+
+**Die Leitplanke für reaktive Fixes** unterscheidet drei Stufen: punktuelle Änderungen (immer
+erlaubt, gewöhnliche Arbeit), korpusweite Änderungen (nur geplant, mit eigenem Issue und
+`work/`-Run - trifft eine Session das Mass-Update-Gate während sie etwas anderes tat, holt sie
+sich nicht den `--confirm`-Token, sondern stoppt und legt ein Issue an) und reaktive Eingriffe
+in Korpusinhalt, um einen Test grün zu machen oder einen Tool-Bug zu umgehen (nie erlaubt,
+Invariante 7). Das Verhältnis zu `kb_dir`/`raw_dir` und `test_pipeline_l0.py` bleibt wie im
+ursprünglichen Befund: kleiner, isolierter Fall in der Fixture, großer, vernetzter Fall in
+`kb/` - keine Fixture-Extraktion aus dem Korpus.
+
+Dev-only und rein additiv - kein Feld, kein Kommando, keine Datei außerhalb von
+`instructions/dev/` ändert sich, daher `--minor` ohne `--breaking`.
+
+**Migration:** none required.
+
+Berührt: `instructions/dev/corpus-policy.md` (neu),
+`instructions/dev/stack-dev/SKILL.md` (Schritt 2, Routing-Zeile).
+
+---
+
+## 4.1.2 - 2026-09-03 - Issue-Abschluss ist ein Body-Rewrite, nicht nur ein Kommentar
+
+**Author:** Torben Nehmer
+
+Aufgefallen beim Schließen von #44: der Abschlussbericht stand als Kommentar da, der Body
+darunter weiterhin als offene Arbeit — Abschnitt „Zu entscheiden" über eine längst getroffene
+Entscheidung, ungehakte Checkliste, Präsens über einen Defekt, den es nicht mehr gab.
+
+Die Regel gab es dafür schon: Schritt 2 von `instructions/dev/issue-tracking.md` sagt, der Body
+ist die aktuelle Wahrheit und wird umgeschrieben, wenn sich der Stand ändert. Nur ließ die
+Formulierung offen, *wann* — und Schritt 7 („Close with what actually happened") war vollständig
+erfüllbar, ohne den Body anzufassen. Ein Abschlussbericht im Kommentar fühlt sich beim Schreiben
+vollständig an; dass der Body dabei zurückbleibt, merkt erst der nächste Leser.
+
+**Schritt 2 ist deshalb schärfer geworden: der Body ist das Plan-File dieses Stacks.** Dasselbe,
+was das Plan-Dokument eines Harness ist, und genauso gepflegt — fortlaufend, sobald etwas darin
+nicht mehr stimmt, nicht am Ende. Der Maßstab ist der Abbruch, nicht der Meilenstein: eine
+Session kann jederzeit enden, und was der Body in diesem Moment sagt, ist die vollständige
+Übergabe. Eine frische Session muss zu **jedem** Zeitpunkt allein aus dem Body weiterarbeiten
+können, ohne Kommentare rückwärts zu lesen und ohne einen Menschen, der es neu erklärt. Entschieden
+ersetzt die Frage, erledigt hakt das Kriterium ab, verworfen steht mit Begründung dort, wo das
+Kriterium stand.
+
+Schritt 7 ist damit kein Sonderakt mehr, sondern die letzte dieser Aktualisierungen: erst Body
+auf den Endstand, dann schließen, dann die Changelog-Zeile aus Schritt 3. Wer Schritt 2 befolgt
+hat, ist fast fertig; wer nicht, zahlt die ganze Schuld im schlechtesten Moment — der
+geschlossene Body ist die Fassung, die danach alle lesen und niemand mehr aufsucht. #44 steht
+als Beispiel drin.
+
+Schritt 3 zieht die Konsequenz: **ein Kommentar pro Session-Umfang, nicht pro Edit.** Ein
+fortlaufend gepflegter Body mit einem Changelog-Kommentar je Änderung wäre Lärm; triviale Pflege
+braucht gar keinen. Der `stack-dev`-Skill sagt es beim Aufgreifen mit, weil dort die Entscheidung
+fällt, ob eine Session den Body überhaupt anfasst.
+
+**Und die ehrliche Antwort auf die Frage nach dem Tooling: es gibt keins, und es soll keins
+geben.** `wikitool` kennt diesen Tracker nicht. Es wird an Instanzen ausgeliefert, die unter
+dieser URL keine Issues haben, während `instructions/dev/` von `dist export` gepruned wird —
+ein Gitea-Client im ausgelieferten Tool wäre eine Dev-Abhängigkeit, die jede Instanz mitträgt,
+um ein Board zu prüfen, das keine von ihnen hat. Der Tracker ist ausschließlich über
+`gitea-mcp` erreichbar, also in einer Session, durch einen Agenten.
+
+Kein `docs verify` fängt also einen geschlossenen Issue, dessen Body offen klingt, einen Body,
+der seinen eigenen Kommentaren widerspricht, oder ein fehlendes Pflichtlabel. Das steht jetzt
+als eigener Abschnitt „What no tool checks" in der Instruktion — nicht als Bedauern, sondern als
+Begründung dafür, warum die Reihenfolge in Schritt 7 ausgeschrieben ist statt aus Schritt 2
+erschlossen zu werden.
+
+---
+
+## 4.1.1 - 2026-09-03 - Testisolation: kb_dir repointet config.ROOT, lint löst Kollektionen gegen den übergebenen Baum auf
+
+**Author:** Torben Nehmer
+
+Issue #44, gefunden beim Bau der Migrations-Gate-Tests für 4.1.0: die `kb_dir`-Fixture baute
+ihren Baum unter `tmp_path`, ließ `config.ROOT` aber auf dem echten Checkout stehen. Jeder
+Codepfad, der eine Datei über `config.ROOT`/`config.KB_DIR` auflöst statt über das übergebene
+Verzeichnis, traf damit das echte Repository.
+
+**Der laute Fall** war ein Test, der `kb_state.write_kb_state()` rief und dabei das
+`.wikitool-kb.json` des Repos überschrieb — Applied-Ledger leer statt zwei Einträgen. In
+`git status` sofort sichtbar und reversibel; bei einer gitignorierten Datei wäre es das nicht
+gewesen.
+
+**Der stillere Fall** ist der teurere. `lint`s Kollektions-Lookup löste eine Seite gegen
+`config.KB_DIR` auf. Für eine Seite unter `tmp_path/kb/` warf das `ValueError`, die Funktion
+antwortete „keine Kollektion", und die Label-Autorisierung übersprang die Kante wortlos.
+`unauthorised_labels` war damit faktisch ungetestet — jeder Test, der das Finding hätte
+auslösen können, bekam eine leere Liste und behauptete nichts. Ein grüner Lauf, der wie eine
+Zusicherung aussah.
+
+**Der Fix ist der Codepfad, nicht die Fixture.** `run_lint()` bekommt ein Verzeichnis
+übergeben und löst jetzt auch intern dagegen auf; `authorised_labels()` bekommt denselben Baum
+gereicht, statt auf `config.KB_DIR` zurückzufallen. Der Regressionstest lintet einen Baum, von
+dem `ROOT` bewusst wegzeigt — genau der Fall, den die alte Auflösung verschluckte. Eine Funktion,
+die ein Verzeichnis entgegennimmt, löst dagegen auf: keine Fixture kann diese Form von außen
+reparieren.
+
+**Beide Korpus-Fixturen repointen jetzt.** `kb_dir` tut, was `raw_dir` längst tat — `ROOT` auf
+das eigene `tmp_path`, plus `use_shipped_type_specs()`. Der Suite-Lauf kippte dadurch keinen
+einzigen Test. Die lokale `rooted_kb`-Umgehung aus 4.1.0 entfällt damit; die Auswahl zwischen
+zwei fast gleichen Fixturen war Wissen, das nirgends stand.
+
+**Und ein Wächter für die ganze Klasse.** `repository_tree_guard` (session-scoped, autouse)
+vergleicht `git status --porcelain` vor und nach dem Lauf und lässt die Suite scheitern, wenn
+sich im Checkout etwas bewegt hat — zwei `git status`-Aufrufe pro Lauf, deshalb per Default an.
+Er vergleicht vorher gegen nachher statt einen sauberen Baum zu verlangen, sagt also nichts über
+die unveröffentlichte Arbeit des Entwicklers. Den Verursacher benennt er nicht;
+`CHEMENU_TREE_GUARD=each` prüft nach jedem Test und tut es. Ohne git oder außerhalb eines
+Repositorys sind beide still.
+
+Was der Wächter nicht sieht: eine Prüfung, die unter Test nichts tut, schreibt keine Datei.
+Dagegen hilft nur ein Test, der das Finding tatsächlich auslöst — der neue tut das.
+
+`instructions/dev/testing-conventions.md` hat dafür einen eigenen Abschnitt („Which tree a test
+writes into"), einen Schritt in der Checkliste und die Regel für neue Fixturen.
+
+---
+
 ## 4.1.0 - 2026-09-03 - Link-Taxonomie: Lint-Findings hart ab kb_version 4.0.0, outbound: an das Type-Spec gebunden, part-of/composition als Inversenpaar
 
 **Author:** Torben Nehmer

VERSION

@@ -1 +1 @@
-4.1.0
+4.2.0

instructions/dev/corpus-policy.md

@@ -0,0 +1,94 @@
+---
+type: types/instruction.md
+name: corpus-policy
+description: What "curated enough" means for kb/ when it is demo and testbed at once, the measurable floors that define it, and what a reactive fix to the corpus may and may not do.
+---
+# Keep kb/ curated enough to develop against, without a second corpus
+
+This instance runs one `kb/` for two purposes at once: a public demo and the testbed this stack
+is developed against. There is deliberately no fixture corpus, no `--with-demo` export, and no
+second repository - see Gitea #28. The corpus's size and shape are set by what targeted
+development needs, not by a synthetic fixture size or a demo aesthetic.
+
+## When to run
+
+- Before judging whether the corpus can exercise a change under development - ranking, index
+  scaling, orphan detection, a new label, a new type-spec.
+- Before a reactive fix touches `kb/` content rather than the failing code - the floors below
+  are what decides whether the fix may proceed as-is.
+- Picking up Gitea #28 or #30, or any issue that references this file.
+
+## The floors
+
+Each is mechanically checkable with an existing `wikitool` command; none needs new tool code.
+A floor exists to keep some class of bug observable, not to describe an aesthetic target - so
+when a session is about to make one of these numbers *worse*, that is the signal to stop and
+think, not a number to defend for its own sake.
+
+| Floor | Check | Why this number |
+|---|---|---|
+| Every page type has ≥1 page | `wikitool search --field type=types/<t>.md` | A type with zero pages means its schema, its collection contract and its lint rules are unexercised |
+| Every declared subtype has ≥1 page | `wikitool search --field <x>_type=<v>` | Same reasoning, one level down - `entity_type`, `concept_type`, `source_type` |
+| ≥5 pages corpus-wide with ≥3 `sources:` entries | one-off script, see below | Provenance fan-in - multiple sources backing one claim - is a real case only a handful of pages exercise; fewer than 5 and a provenance-index bug can hide |
+| Orphan pages (no inbound link) between 1 and 10 | `wikitool lint` | Zero orphans makes orphan detection itself unobservable; more than 10 means the corpus stopped being curated |
+| Average outbound wikilinks per page ≥4 | one-off script, see below | Below this, ranking and graph-traversal work has too little structure to exercise |
+
+A floor is a lower bound only. There is no upper bound on page count or on any of these numbers
+except the orphan ceiling above - a corpus that outgrows these floors through real ingests is
+not a problem this file cares about.
+
+**Measured 2026-09-03** (see Gitea #28): 181 pages, 14/14 types and subtypes covered, 12 pages
+with ≥3 sources, 3 orphans, 6.2 average outbound links. All floors held without any manufactured
+content - the corpus was already big enough when the question was asked.
+
+A type or subtype sitting at exactly the floor - one page - shows no set-level bugs, only that
+the type is *reachable*. That is a soft target for the next `wiki-ingest` that happens to
+produce a matching page, never a reason to write one: filing an unsourced page to clear a floor
+is exactly what AGENTS.md invariant 3 forbids, floor or no floor. The same holds for an
+authorised link label with zero live uses (`wikitool xref` reports these) - fill it when a real
+edge calls for it, never manufacture one to exercise the label.
+
+To check the two floors without a dedicated command, walk `kb/**/*.md` (excluding
+`INDEX.md`/`COLLECTION.md`/`CONTRACT.md`/`CONVENTIONS.md`), parse frontmatter, and: count pages
+whose `related:` array (resolved against page titles) has ≥3 entries for outbound density; count
+`sources:` array length ≥3 for the provenance floor. `wikitool search` and `wikitool lint`
+cover everything else in the table.
+
+## What a reactive fix may do to kb/ content
+
+Three tiers, by how much of the corpus a change touches:
+
+1. **Pointwise - always allowed.** Creating, updating, renaming or deleting a single page
+   through the normal tools (`new`, `touch`, the page-lifecycle procedure), below the
+   Mass-Update Gate's threshold. This is ordinary work and needs no special permission.
+2. **Corpus-wide - planned only, never reactive.** A migration, a vocabulary sweep, a bulk
+   `touch` across many pages. This needs its own issue and, per `work/CONTRACT.md`, a `work/`
+   run - never a same-session reaction to whatever the session was originally doing. If a
+   session hits the Mass-Update Gate (exit 42, see `instructions/gates.md`) while working on
+   something else, it does not fetch the `--confirm` token to push through: it stops, opens an
+   issue for the corpus-wide change, and finishes the original task without it.
+3. **Reactive - never allowed.** Deleting or reshaping a page to make a failing test pass;
+   restructuring corpus content to route around a tool bug (AGENTS.md invariant 7); using
+   `kb/` as a scratch surface for a tool experiment. If a stack change under development needs a
+   corpus shape that does not exist, build it as a pytest fixture (see the next section) -
+   never manufacture it in `kb/`.
+
+## Relationship to the test fixtures
+
+`tools/chemenu/tests/conftest.py`'s `kb_dir`/`raw_dir` fixtures and `test_pipeline_l0.py` cover
+the **small, isolated** case: a handful of pages, built fresh per test, hermetic. `kb/` covers
+the **large, connected** case: 181+ pages, grown link density, real provenance history that no
+per-test fixture reconstructs economically. The cut: if a `tmp_path` tree can reproduce what the
+test needs, it belongs in a fixture; if the test needs density or scale that only a grown corpus
+has, it belongs against `kb/`. Neither absorbs the other's job - see
+[testing-conventions.md](testing-conventions.md).
+
+## Decision points
+
+- **A floor would be violated by an in-progress change - is that a blocker?** Only for the
+  orphan ceiling and the type/subtype floors, since those two can go to zero. The density and
+  provenance floors move gradually with ordinary ingests and are not gating on any single
+  session.
+- **Corpus is "too small" for a feature under development?** That is not this file's problem to
+  solve by adding pages - see tier 3 above. Either the feature waits for a real ingest to supply
+  the shape, or it gets a pytest fixture.

instructions/dev/issue-tracking.md

@@ -28,8 +28,12 @@ issues at that URL, which is exactly why `dist export` excludes
   an assumption nobody has checked, a decision that needs the user.
 - Picking an issue up: before doing anything else, read the body as the current
   spec, and re-label it if the ground has moved since.
+- **While working on one:** the body is updated as the state moves, not at the
+  end (step 2). A session that is interrupted leaves the body as its handover.
 - Prioritising: deciding what to pick up next, or re-labelling after the ground
   moved.
+- Closing one: the body is rewritten to its final state first, and only then
+  closed (step 7).
 
 ## Steps
 
@@ -38,23 +42,47 @@ issues at that URL, which is exactly why `dist export` excludes
    specific files or commands involved. An issue that only makes sense to
    whoever wrote it is a note, and notes were the problem.
 
-2. **Treat the body as the current truth, not as a historical first post.**
-   Work on one issue spans several sessions, often weeks apart, and the body is
-   the only thing that connects them: a session opening the issue must be able
-   to reconstruct what is decided and what is still open from the body alone,
-   without a human re-explaining it. So when the state changes, **rewrite the
-   body** - do not append to a text that has become wrong. An additively grown
-   log forces every later reader to reconstruct the current state by filtering
-   the whole history.
+2. **The body is the working state, not a historical first post - keep it
+   current as you go.** It is this stack's plan file: the same thing a harness's
+   own plan document is, and it is maintained the same way. Not written once,
+   not brought up to date at the end, but **updated whenever something in it
+   stops being true** - a decision made, a criterion met, an approach ruled out,
+   a new constraint found.
+
+   The test is an abort, not a milestone. A session can end at any moment - an
+   interrupt, a context limit, a crash, a human walking away - and whatever the
+   body says at that instant is the entire handover. So the standard is: **at
+   every point, a fresh session must be able to open the body and pick the work
+   up from there**, without a human re-explaining it and without reading back
+   through the comments. If the body would mislead someone who read it right
+   now, it is already out of date, whether or not the work is finished.
+
+   That means updating *during* the work, not only at its end:
+
+   - a decision gets made → the decision and its reasoning replace the question
+   - an acceptance criterion is done → tick it, in the same session that did it
+   - something turns out differently than the issue assumed → the assumption is
+     corrected where it stands, not contradicted three paragraphs later
+   - work is deferred or dropped → say so, with the reason, where the criterion is
+
+   **Rewrite, never append.** Do not add to a text that has become wrong: an
+   additively grown log forces every later reader to reconstruct the current
+   state by filtering the whole history, which is the exact cost the body exists
+   to remove. Comments carry the history (step 3); the body carries the state.
 
    Body rewrites and comments are an LLM session's job. A human normally
    touches only labels and metadata directly.
 
-3. **Comment a changelog, never a copy.** Every body rewrite gets one short
-   comment naming only what changed against the previous state - what is new,
-   what is gone, what was corrected. Do not snapshot the old body into a
-   comment: a full copy per revision forces a human to diff two prose texts,
-   which is not a readable history, only another copy.
+3. **Comment a changelog, never a copy.** A body rewrite gets one short comment
+   naming only what changed against the previous state - what is new, what is
+   gone, what was corrected. Do not snapshot the old body into a comment: a full
+   copy per revision forces a human to diff two prose texts, which is not a
+   readable history, only another copy.
+
+   One comment per *session's worth* of change, not per edit. Step 2 asks the
+   body to be kept current continuously, and a comment for every tick would bury
+   the board in noise; the changelog line summarises what that session moved.
+   Trivial upkeep - a typo, a tightened sentence - needs no comment at all.
 
    ```
    **Changelog:** Decision 2 tightened - `kind/` may now change over an
@@ -127,10 +155,46 @@ issues at that URL, which is exactly why `dist export` excludes
    answered can drop a size and move `kind/decision` to `kind/build`. Silent
    re-labelling is how a board stops meaning anything.
 
-7. **Close with what actually happened**, not with a commit hash alone: which
-   proposals were implemented, which were deliberately left out and why, and
-   what was verified. The issue is the only place that record survives - a
-   changelog entry says what changed, not what was decided against.
+7. **Closing is the last body update, not a comment.** If step 2 was followed
+   the body is already nearly there, and closing only settles what the final
+   run established. If it was not, closing is where the whole debt comes due -
+   and it comes due at the worst moment, because a closed body is the version
+   everyone reads afterwards and nobody revisits.
+
+   Either way the body reaches its final state *before* the issue closes:
+   proposals that were decided read as decided, a "to decide" section has become
+   the decision with its reasoning, acceptance criteria are ticked or struck with
+   a reason, and what was verified is named. Then close, with the one-line
+   changelog comment step 3 asks for.
+
+   Record what actually happened, not a commit hash alone: which proposals were
+   implemented, which were deliberately left out and why, and what was verified.
+   The issue is the only place that record survives - a changelog entry says
+   what changed, not what was decided against.
+
+   **A closing report in a comment does not satisfy this.** It reads as
+   complete to whoever writes it and leaves a body still phrased as open work:
+   unticked boxes, an undecided decision section, present tense about a defect
+   that no longer exists. #44 closed exactly that way, with a thorough comment
+   above a body that still asked for a decision that had already been made and
+   shipped. Nothing mechanical catches it (see below), which is why it is a step
+   rather than a habit.
+
+## What no tool checks
+
+`wikitool` does not know this tracker exists, and should not learn. It ships to
+instances that have no issues at that URL, while this file and the workflow it
+describes are pruned by `dist export` - a Gitea client inside the shipped tool
+would be a dev-only dependency carried by every instance, to check a board none
+of them have. The tracker is reachable only through the `gitea-mcp` server, in a
+session, by an agent.
+
+So there is no `docs verify` for the board. Nothing reports a closed issue whose
+body still reads as open, a body that contradicts its own comments, or an issue
+missing one of the four mandatory labels. Every one of those is caught by a
+session following this file, or not at all - which is the argument for the
+sequence in step 7 being explicit about the order (body first, then close),
+rather than leaving it to be inferred from step 2.
 
 ## Decision points
 
@@ -144,7 +208,8 @@ issues at that URL, which is exactly why `dist export` excludes
 - **Rewrite the body, or add a comment?** Rewrite whenever a reader of the body
   alone would otherwise be misled - a changed decision, a dropped criterion, a
   new constraint. A comment carries the changelog line for that rewrite, and
-  nothing else that a future session needs in order to act.
+  nothing else that a future session needs in order to act. Closing an issue is
+  always a rewrite - see step 7.
 - **An old issue carries only `prio/` and `size/`?** Complete it to all four
   when you touch it, rather than in a sweep. The board reaches the new scheme
   issue by issue, as each is picked up.

instructions/dev/stack-dev/SKILL.md

@@ -43,15 +43,21 @@ stack development happens in the origin repo instead (see AGENTS.md's routing li
    engineering, memory and deploy-time learning; consult before a design decision in those
    areas.
    [issue-tracking.md](../issue-tracking.md) - open work lives in Gitea issues, one per work
-   package, labelled `area/`, `kind/`, `prio/` and `size/`, with the body kept as the current
-   truth rather than as a first post. There is no `TODO.md`. Read it before filing something
-   for later, before editing an issue, or before deciding what to pick up next.
+   package, labelled `area/`, `kind/`, `prio/` and `size/`. There is no `TODO.md`. **The body
+   of the issue you are working on is this session's plan file:** keep it current as the state
+   moves, not at the end, so an interrupted session leaves a body the next one can resume from.
+   Read it before filing something for later, before editing or closing an issue, or before
+   deciding what to pick up next.
    [testing-conventions.md](../testing-conventions.md) - the suite runs against a deliberately
    empty machine; what the autouse fixture already neutralizes, and what a test still has to
    establish itself. Read it before adding or changing a test.
    [version-parts.md](../version-parts.md) - which part a change bumps: the drop-in test, the
    catalogue of breaks that cross the compatibility boundary with `kb/` untouched, and what to
    put in front of the user before a breaking bump. Read it before step 3.
+   [corpus-policy.md](../corpus-policy.md) - what "curated enough" means for the shared
+   demo/testbed `kb/`, the measurable floors that define it, and what a reactive fix may and may
+   not do to corpus content. Read it before judging whether the corpus can exercise a change, or
+   before any fix that would touch `kb/` content.
    More instructions are added here incrementally as stack-development needs come up - this
    list grows without needing this skill file to change shape.
 3. **Raise the version, if the change ships.** A change under `tools/`, `types/`,

instructions/dev/testing-conventions.md

@@ -40,6 +40,48 @@ resolved paths and `conventions`' parsed `kb/CONVENTIONS.md`. A test that *rewri
 conventions file mid-test calls `conventions.reset_cache()` itself - the fixture answers for the
 boundary between tests, not for one inside a test.
 
+## Which tree a test writes into
+
+The environment is one half of the isolation; `config.ROOT` is the other. With `CHEMENU_ROOT`
+cleared, `ROOT` falls back to the checkout pytest is running from - deliberately, because most
+tests want the shipped `types/`. It also means that any code path resolving a file through
+`config.ROOT` or `config.KB_DIR` reaches **the real repository**, no matter which tree the
+fixture built.
+
+Both corpus fixtures therefore repoint it: `raw_dir` and `kb_dir` each set
+`config.ROOT` to their `tmp_path` and re-declare the shipped `types/` through
+`use_shipped_type_specs()`. `config`'s module `__getattr__` resolves the derived paths on
+access, so repointing `ROOT` carries `KB_DIR`, `RAW_DIR` and the rest with it. A new fixture
+that builds a tree does the same thing - that is the rule here, not a per-test judgment.
+
+`kb_dir` did not, until Gitea #44. Two things came of that. A test calling
+`kb_state.write_kb_state()` overwrote the real `.wikitool-kb.json`, which `git status` made
+visible within the minute. Quieter and worse: `lint`'s collection lookup resolved a page
+against `config.KB_DIR`, so every fixture page read back as "no collection" and the
+`unauthorised_labels` check skipped every edge in silence - the finding had no working test at
+all, and its green run read like an assurance.
+
+Two guards came out of it, both in `conftest.py`:
+
+| Guard | Default | Cost |
+|---|---|---|
+| `repository_tree_guard` (session) | on | two `git status --porcelain` calls per run |
+| `per_test_tree_guard` | off, `CHEMENU_TREE_GUARD=each` turns it on | one `git status` per test |
+
+The session guard compares the working tree before against after and fails the run if anything
+moved, so it says nothing about uncommitted work a developer already had. It cannot name the
+test that did it; `CHEMENU_TREE_GUARD=each` can, and is the way to bisect once it fires. Where
+git is unavailable or the checkout is not a repository, both are silently inert.
+
+Neither guard sees the second, quieter half: a check that silently *does nothing* under test
+writes no file. That one is only caught by a test that asserts the finding actually fires -
+which is why `test_unauthorised_label_is_judged_in_a_tree_that_is_not_the_configured_kb`
+lints a tree `ROOT` deliberately points away from.
+
+**A function that takes a directory resolves against that directory.** `run_lint(kb_dir)`
+reading `config.KB_DIR` for one of its own lookups was the defect behind the quiet half, and
+no fixture can fix that shape from the outside.
+
 ## When to run
 
 Whenever you add or change a test under `tools/chemenu/tests/`.
@@ -80,7 +122,12 @@ Whenever you add or change a test under `tools/chemenu/tests/`.
    `conftest.py` in the same change. A variable the tool reads and the fixture does not clear
    is the exact hole this whole file is about, reopened.
 
-5. **Verify against an empty machine before publishing**, not only in your own shell:
+5. **Writing a fixture that builds a tree?** Repoint `config.ROOT` at it and call
+   `use_shipped_type_specs(monkeypatch)`, as `raw_dir` and `kb_dir` do - see
+   [Which tree a test writes into](#which-tree-a-test-writes-into). A fixture that returns a
+   path without repointing hands the code under test the real repository.
+
+6. **Verify against an empty machine before publishing**, not only in your own shell:
 
    ```bash
    cd tools && env -i PATH="$PATH" HOME="$(mktemp -d)" \
@@ -92,7 +139,7 @@ Whenever you add or change a test under `tools/chemenu/tests/`.
    `.venv/bin/python -m pytest -q`. A difference between the two is a leak, and the leaking
    variable belongs in step 4's list.
 
-6. **Check the coverage report when adding tests to close a gap**, rather than guessing which
+7. **Check the coverage report when adding tests to close a gap**, rather than guessing which
    lines were uncovered:
 
    ```bash

tools/chemenu/lint_core.py

@@ -165,9 +165,15 @@ def run_lint(kb_dir: Path) -> dict:
     # propagated, a deleted page, or a URL pasted where a title belongs - used
     # to pass every check. Which fields hold page titles is declared by each
     # type-spec's `page_ref_fields:`, not hardcoded here.
+    # Resolved against the directory `run_lint()` was handed, not against
+    # `config.KB_DIR`. A page under a tree that is not the configured corpus -
+    # every fixture tree, and any `lint <path>` aimed elsewhere - raised
+    # `ValueError` here and read as "no collection", which made the label
+    # authorisation below skip the edge in silence rather than judge it
+    # (Gitea #44).
     def _collection_of(page):
         try:
-            return page.path.relative_to(config.KB_DIR).parts[0]
+            return page.path.relative_to(kb_dir).parts[0]
         except (ValueError, IndexError):
             return None
 
@@ -211,7 +217,9 @@ def run_lint(kb_dir: Path) -> dict:
                 destination = _collection_of(target_page)
                 if destination is None:
                     continue
-                allowed = kb_collections.authorised_labels(source_collection, destination)
+                allowed = kb_collections.authorised_labels(
+                    source_collection, destination, kb_dir
+                )
                 if edge.label not in allowed:
                     unauthorised_labels.append(
                         {

tools/chemenu/tests/conftest.py

@@ -1,4 +1,5 @@
 import os
+import subprocess
 from pathlib import Path
 
 import pytest
@@ -37,6 +38,78 @@ _GIT_ENV = (
 )
 
 
+def _working_tree_state() -> str | None:
+    """`git status --porcelain` for the checkout the tests live in, or None if
+    there is no git available to ask."""
+    try:
+        result = subprocess.run(
+            ["git", "-C", str(config._PACKAGE_ROOT), "status", "--porcelain"],
+            capture_output=True,
+            text=True,
+            timeout=60,
+        )
+    except (OSError, subprocess.SubprocessError):
+        return None
+    return result.stdout if result.returncode == 0 else None
+
+
+_TREE_GUARD_MESSAGE = (
+    "A test wrote into the repository checkout instead of into its tmp_path.\n"
+    "`git status --porcelain` moved while the suite ran:\n\n"
+    "  before:\n{before}\n"
+    "  after:\n{after}\n\n"
+    "This is the class of bug Gitea #44 describes: code under test resolves a "
+    "path through `config.ROOT`/`config.KB_DIR` rather than through the "
+    "directory the fixture handed it, so the write lands in the real tree. Fix "
+    "the fixture (repoint `config.ROOT`, as `raw_dir` and `kb_dir` do) or the "
+    "code path, never the symptom.\n"
+    "To find the test that did it, re-run with CHEMENU_TREE_GUARD=each - the "
+    "guard then checks after every test and fails on the first one that moves "
+    "the tree."
+)
+
+
+@pytest.fixture(scope="session", autouse=True)
+def repository_tree_guard():
+    """Fail the run if the suite moved a file in the real checkout.
+
+    Two `git status` calls for the whole session, which is why this is on by
+    default: it catches the whole class rather than the one case that was
+    noticed. It compares before against after rather than demanding a clean
+    tree, so it says nothing about a developer's own uncommitted work.
+
+    It cannot name the culprit - set `CHEMENU_TREE_GUARD=each` for that, which
+    trades a `git status` per test for a failure on the test that did it.
+    """
+    before = _working_tree_state()
+    yield
+    after = _working_tree_state()
+    if before is None or after is None or before == after:
+        return
+    raise AssertionError(
+        _TREE_GUARD_MESSAGE.format(before=before or "(clean)", after=after or "(clean)")
+    )
+
+
+@pytest.fixture(autouse=True)
+def per_test_tree_guard(repository_tree_guard):
+    """The bisect half of `repository_tree_guard`, off unless asked for.
+
+    `CHEMENU_TREE_GUARD=each` turns the session-wide "something moved the tree"
+    into "this test moved the tree", at the cost of a `git status` per test.
+    """
+    if os.environ.get("CHEMENU_TREE_GUARD") != "each":
+        yield
+        return
+    before = _working_tree_state()
+    yield
+    after = _working_tree_state()
+    if before is not None and after is not None and before != after:
+        raise AssertionError(
+            _TREE_GUARD_MESSAGE.format(before=before or "(clean)", after=after or "(clean)")
+        )
+
+
 @pytest.fixture(autouse=True)
 def hermetic_environment(tmp_path: Path, monkeypatch: pytest.MonkeyPatch):
     """Cut every test off from the machine it runs on.
@@ -159,13 +232,25 @@ def raw_dir(tmp_path: Path, monkeypatch: pytest.MonkeyPatch) -> Path:
 
 
 @pytest.fixture
-def kb_dir(tmp_path: Path) -> Path:
+def kb_dir(tmp_path: Path, monkeypatch: pytest.MonkeyPatch) -> Path:
     """A minimal fixture kb/ with the standard collection layout, populated
     with a handful of pages covering entities/concepts/sources/comparisons.
 
     Every collection carries a COLLECTION.md, both because that is what makes it
     a collection and because the scanner must prove it skips them at a depth the
-    kb-root meta files never reach."""
+    kb-root meta files never reach.
+
+    `config.ROOT` is repointed for the same reason `raw_dir` does it, one
+    collection over: code under test that resolves a path through
+    `config.ROOT`/`config.KB_DIR` rather than through the directory it was
+    handed otherwise reaches the *real* repository. That was not theoretical
+    either - a test calling `kb_state.write_kb_state()` overwrote this
+    checkout's `.wikitool-kb.json`, and `lint`'s collection lookup answered
+    "no collection" for every fixture page, which left `unauthorised_labels`
+    with no working test at all (Gitea #44).
+    """
+    monkeypatch.setattr(config, "ROOT", tmp_path)
+    use_shipped_type_specs(monkeypatch)
     kb = tmp_path / "kb"
     for sub in ("entities/projects", "entities/systems", "entities/tools",
                 "entities/technologies", "entities/people",

tools/chemenu/tests/test_hermetic_env.py

@@ -74,3 +74,23 @@ def test_wiki_author_overrides_the_git_identity(tmp_path: Path,
     monkeypatch.setattr(config, "ROOT", tmp_path)
     monkeypatch.setenv("WIKI_AUTHOR", "Env Override")
     assert config.default_author() == "Env Override"
+
+
+def test_kb_dir_repoints_the_configured_root_at_its_own_tree(kb_dir: Path, tmp_path: Path):
+    """The other half of the isolation, and the one `kb_dir` was missing until
+    Gitea #44: a fixture that builds a corpus but leaves `config.ROOT` on the
+    real checkout hands every `config.KB_DIR` lookup the developer's own wiki -
+    which is how a test overwrote the repository's `.wikitool-kb.json`."""
+    assert config.ROOT == tmp_path
+    assert config.KB_DIR == kb_dir
+
+
+def test_raw_dir_repoints_the_configured_root_at_its_own_tree(raw_dir: Path, tmp_path: Path):
+    assert config.ROOT == tmp_path
+    assert config.RAW_DIR == raw_dir
+
+
+def test_both_corpus_fixtures_keep_the_shipped_type_specs_reachable(kb_dir: Path):
+    """Repointing `ROOT` moves `TYPES_DIR` with it, so the repoint has to be
+    paired with `use_shipped_type_specs()` or no page type resolves at all."""
+    assert (config.TYPES_DIR / "entity.md").is_file()

tools/chemenu/tests/test_lint.py

@@ -15,7 +15,6 @@ from chemenu.commands.lint import (
 )
 from chemenu.frontmatter_io import write_page
 from chemenu.provenance import cite_id, render_cite_block
-from chemenu.tests.conftest import use_shipped_type_specs
 from chemenu.version import Version
 
 
@@ -482,26 +481,18 @@ def test_lint_does_not_count_a_shell_prompt_as_a_quote(kb_dir):
 
 # --- the migration gate on `unlabelled_edges` / `unauthorised_labels` -------
 #
-# These four repoint `config.ROOT` at the fixture tree, which the `kb_dir`
-# fixture alone does not do. Two things need it: `write_kb_state()` writes
-# `.wikitool-kb.json` relative to `ROOT`, and `lint`'s collection lookup
-# resolves a page against `config.KB_DIR` rather than against the directory it
-# was handed - so without the repoint the gate would be read off the real
-# repository's state file.
+# These read and write `.wikitool-kb.json`, which `kb_state` resolves relative
+# to `config.ROOT`. The `kb_dir` fixture repoints `ROOT` at its own tmp_path
+# (Gitea #44), so the gate is read off the fixture tree; before it did, these
+# four ran against the real repository's state file and one of them overwrote
+# it.
 
 
-@pytest.fixture
-def rooted_kb(kb_dir, tmp_path, monkeypatch):
-    monkeypatch.setattr(config, "ROOT", tmp_path)
-    use_shipped_type_specs(monkeypatch)
-    return kb_dir
-
-
-def _page_with_an_unlabelled_edge(rooted_kb):
+def _page_with_an_unlabelled_edge(kb_dir):
     """A `related:` entry that is a bare title rather than a `label: title`
     mapping - the shape every page was in before the 4.0.0 migration."""
     write_page(
-        rooted_kb / "entities/tools/bare-edge.md",
+        kb_dir / "entities/tools/bare-edge.md",
         {"type": "types/entity.md", "entity_type": "tool", "tags": [], "created": "2026-09-03",
          "modified": "2026-09-03", "related": ["Modbus"], "sources": [], "confidence": 0.8,
          "provenance": "general", "summary": "One edge whose label was never declared."},
@@ -509,14 +500,14 @@ def _page_with_an_unlabelled_edge(rooted_kb):
     )
 
 
-def test_unlabelled_edge_is_advisory_below_kb_version_4(rooted_kb):
+def test_unlabelled_edge_is_advisory_below_kb_version_4(kb_dir):
     """The window the migration document describes: the machinery has landed,
     the corpus has not been converted yet, and `lint --fail-on-error` must not
     refuse the very tree the migration tells the instance to publish unit by
     unit."""
-    _page_with_an_unlabelled_edge(rooted_kb)
+    _page_with_an_unlabelled_edge(kb_dir)
     kb_state.write_kb_state(Version(3, 0, 0), [])
-    report = run_lint(rooted_kb)
+    report = run_lint(kb_dir)
     assert report["unlabelled_edges"] != []
     assert "unlabelled_edges" not in hard_error_keys()
     # Narrowed to the finding under test: the fixture corpus carries unrelated
@@ -525,34 +516,64 @@ def test_unlabelled_edge_is_advisory_below_kb_version_4(rooted_kb):
     assert has_hard_errors({"unlabelled_edges": report["unlabelled_edges"]}) is False
 
 
-def test_unlabelled_edge_is_hard_at_kb_version_4(rooted_kb):
+def test_unlabelled_edge_is_hard_at_kb_version_4(kb_dir):
     """Once the migration is recorded, a bare title is no longer a page waiting
     its turn - it is an edge whose author did not say what it asserts."""
-    _page_with_an_unlabelled_edge(rooted_kb)
+    _page_with_an_unlabelled_edge(kb_dir)
     kb_state.write_kb_state(Version(4, 0, 0), [])
-    report = run_lint(rooted_kb)
+    report = run_lint(kb_dir)
     assert report["unlabelled_edges"] != []
     assert "unlabelled_edges" in hard_error_keys()
     assert has_hard_errors({"unlabelled_edges": report["unlabelled_edges"]}) is True
 
 
-def test_unauthorised_label_is_hard_at_kb_version_4(rooted_kb):
+def test_unauthorised_label_is_hard_at_kb_version_4(kb_dir):
     """The fixture contracts authorise `depends-on` but not `contradicts`."""
     write_page(
-        rooted_kb / "entities/tools/off-menu.md",
+        kb_dir / "entities/tools/off-menu.md",
         {"type": "types/entity.md", "entity_type": "tool", "tags": [], "created": "2026-09-03",
          "modified": "2026-09-03", "related": [{"contradicts": "Modbus"}], "sources": [],
          "confidence": 0.8, "provenance": "general", "summary": "A label off this menu."},
         "\n# off-menu\n\nA label the source collection never authorised.\n",
     )
     kb_state.write_kb_state(Version(4, 0, 0), [])
-    report = run_lint(rooted_kb)
+    report = run_lint(kb_dir)
     assert report["unauthorised_labels"] != []
     assert "unauthorised_labels" in hard_error_keys()
     assert has_hard_errors({"unauthorised_labels": report["unauthorised_labels"]}) is True
 
 
-def test_a_tree_that_never_declared_a_kb_version_keeps_every_key(rooted_kb):
+def test_unauthorised_label_is_judged_in_a_tree_that_is_not_the_configured_kb(
+    kb_dir, tmp_path, monkeypatch
+):
+    """`run_lint()` judges the tree it was handed, not the configured corpus.
+
+    The collection lookup used to resolve a page against `config.KB_DIR`; a
+    page anywhere else raised `ValueError`, read back as "no collection", and
+    the label check skipped the edge without a word. That is why
+    `unauthorised_labels` was untested in practice before Gitea #44 - every
+    fixture tree was somewhere else. Here `ROOT` deliberately points away from
+    the tree under lint, which is the case the old code got wrong.
+    """
+    write_page(
+        kb_dir / "entities/tools/off-menu.md",
+        {"type": "types/entity.md", "entity_type": "tool", "tags": [], "created": "2026-09-03",
+         "modified": "2026-09-03", "related": [{"contradicts": "Modbus"}], "sources": [],
+         "confidence": 0.8, "provenance": "general", "summary": "A label off this menu."},
+        "\n# off-menu\n\nA label the source collection never authorised.\n",
+    )
+    elsewhere = tmp_path / "elsewhere"
+    elsewhere.mkdir()
+    monkeypatch.setattr(config, "ROOT", elsewhere)
+    assert config.KB_DIR != kb_dir
+
+    report = run_lint(kb_dir)
+    assert {
+        "page": "off-menu", "target": "Modbus", "label": "contradicts", "destination": "concepts"
+    } in report["unauthorised_labels"]
+
+
+def test_a_tree_that_never_declared_a_kb_version_keeps_every_key(kb_dir):
     """No `.wikitool-kb.json` means a fresh instance, which starts at the
     current shape rather than migrating into it - so there is no outstanding
     migration for a gated finding to be the noise of."""